Invariant representations for action recognition in the visual system